Legal Document Preparers: Intake Forms and Assembly With AI

A legal document preparer can use AI to turn intake answers into clean structured data, flag blanks and contradictions, and fill approved templates. It must not choose forms for the customer, interpret the law, or suggest strategy. Keep AI on clerical work, keep the choices with the customer, and have a person check every page before it leaves.

The rules on what a non-lawyer preparer may do are set locally and differ from place to place, so treat what follows as a working method, not a legal opinion. Read your own licence conditions or regulator guidance first. If anything in your scope is unclear, ask a lawyer before you automate it, because automating an unclear step only makes the mistake faster.

Follow me on Instagram@sagnikteaches

The line your tool must never cross

Preparer rules vary, but most circle one idea. One court-run certification programme puts it in a sentence: certified preparers may provide general legal information but may not give legal advice. In practice that usually means you may help a person complete documents they have chosen, using information they supply; you may not advise on which documents suit their situation, how the law applies to their facts, or what they should do. An AI chat tool will happily do all three if asked, which is why the boundary has to be built into your process and not left to good intentions.

Connect on LinkedInSagnik Bhattacharya

When customers search for this topic they usually want to know what a preparer is. Say it in the first minute of your intake: you prepare paperwork from the customer's answers, you are not a lawyer, and you cannot tell them what to file. That one paragraph, written in your own words and agreed with your regulator's guidance, does more to protect you than any tool setting.

Subscribe on YouTube@sagnikteaches

Step one: design the intake form around fields, not paragraphs

An AI assembly flow is only as good as the data going in. The common mistake is a long free-text box that says "tell us about your situation", which invites exactly the narrative that drifts into advice territory. Build intake as short, typed fields, with a separate box for anything the customer wants to add.

Say a preparer who helps customers complete routine court paperwork uses this structure, illustrative and trimmed:

1. Who is asking: full legal name as on ID, any other names used, date of birth, address, phone, email, preferred contact method.
2. Other people involved: names, relationship, addresses where the customer knows them, and a tick for "I do not know".
3. The paperwork: which document the customer says they want prepared, in their words, plus a field for "who told you to file this" (court, lawyer, agency, nobody).
4. Key dates: only dates printed on papers the customer holds, with a photo upload of the paper.
5. Money figures: amounts from documents, with a tick for "estimate".
6. Anything urgent: hearing within 14 days, safety concerns, children's arrangements, a deadline printed on a notice.
7. Consents: data processing, use of an AI tool, storage period.

The "who told you to file this" question is quietly useful. It records where the customer's choice came from, which is a good file note if anyone ever asks whether you steered them. The "urgent" box feeds your red-flag rule in step five.

Step two: turn the answers into a record with AI

This is where AI earns its place. Customers type "mum's address is the one in Elm Road, number 14, flat b" and "got the letter 3rd march". A tool can turn this into consistent data. The prompt below keeps the model in a clerical lane. It sees a customer code, not a name, and is forbidden from adding facts.

You are a data-entry assistant for a legal document preparer. You never give legal
advice and never suggest what the customer should file.

Task: convert the intake answers below into JSON with these keys only:
customer_code, legal_name, other_names, dob, address, other_parties (list of
{name, relationship, address}), document_requested, document_source,
key_dates (list of {label, date_iso, source_document}), money (list of
{label, amount, is_estimate}), urgent_flags (list), missing_or_unclear (list).

Rules: use only what is written. If a value is absent or ambiguous, leave it null and
add a line to missing_or_unclear. Convert dates to YYYY-MM-DD. Do not correct names.
Do not comment on the law.

INTAKE:
Customer code: C-2217
Name: [legal first name] A. [surname] (signs some letters with a nickname)
Other party: the landlord, [landlord name], 14 Elm Road flat b
Wants: "the response form for the letter"
Letter from: the court, dated 3rd march, says reply within 14 days
Rent figure on the letter: 1,150 a month, "think I owe about 3 months"

A realistic, labelled-as-illustrative reply: document_requested is "the response form for the letter", document_source is "the court", the key date is 2026-03-03 labelled "date of letter", the 14-day reply window goes into urgent_flags, the rent is 1,150 with is_estimate false, the arrears are 3 months with is_estimate true, and missing_or_unclear contains "year of the letter not stated" and "which specific form the letter refers to". Notice the last item: the tool has not guessed the form. It has handed you a question for the customer, which you ask without telling them what the answer should be.

What you would fix: the model may slip the nickname into the legal name field if the prompt is loose, which is why "do not correct names" is in the rules. Always read the JSON against the customer's own paperwork. Two minutes of checking here saves a returned filing later.

Step three: assemble from approved templates only

Assembly means taking the clean record and merging it into a template you have already approved: a court's own blank form, a standard letter, or a document from a reputable form publisher. The merge itself needs no language model; a mail-merge, a document-assembly tool or a spreadsheet with a script does it more predictably. Use AI to check, not to write the legal words.

• Template library: one folder per form, with the source and the date you last checked it against the court's current version. Out-of-date forms are the commonest reason filings bounce.
• Mapping sheet: a table listing each form field and the intake field that feeds it. Build it once per form.
• Merge: run the merge from the JSON so one record fills every form the customer asked for.
• AI check: give the model the filled fields as text and ask for inconsistencies only.

Compare the two lists below. Report only: (1) fields that are blank but marked
required, (2) names or dates that differ between forms, (3) amounts that do not add up,
(4) anything formatted differently from the court's example. Do not suggest legal changes.
If everything matches, say "No inconsistencies found."

FORM A FIELDS: ...
FORM B FIELDS: ...
INTAKE RECORD: ...

A sample reply might say that the respondent's address on form B differs from the intake record by one character, and that the arrears total of 3,450 assumes exactly three unpaid months at 1,150, although the customer described "about 3 months" as an estimate and the record says nothing about part payments. The second item is a question to put back to the customer, not a correction to make. That is the difference between a check and advice.

Step four: plain-language explanations from your own text

Customers often cannot read the boilerplate on a form. You may be allowed to explain what a box is asking for, as printed on the form's own instructions, but not what the customer should write. Keep a short approved help text for each field and let AI only rephrase it at a reading age you choose.

Before (approved help text): "Enter the total sum, if any, you acknowledge as due to the claimant as at the date of this form."

After (plain rewrite, still no advice): "This box asks for the amount of money, if any, that you agree you owe the person who started the case. It is your own figure. If you are not sure, you can leave it and speak to a lawyer or a free advice service before you sign."

The last sentence is the safety valve: every explanation points to professional help for anything beyond the form's wording. Review the rewrite once, store it as the new approved text, and reuse it. Do not regenerate it per customer, because a fresh generation each time means a fresh chance of drift.

A mistake that crosses the line without anyone noticing

Here is a realistic slip. A preparer, short of time, pastes a customer's whole story into a general chat tool and types "what is the best form for this?". The reply names a form and adds, helpfully, "you should also claim for costs". The preparer prints the form. In one step the tool has chosen the document and suggested a tactic, and the preparer has acted on both. No one intended advice, and it still happened.

Three controls prevent it. First, the only prompts staff may use are the stored ones above, in a shared document. Second, the tool is configured, or the workspace instructions say, that it refuses legal questions and redirects to "ask a lawyer". Third, the review step asks the reviewer to confirm that the form came from the customer's stated choice, with the field in the record to prove it.

A second customer: a signature-light agency form

Not every job is a court deadline. Say a customer wants help completing a straightforward agency application about a change of name on records. The intake answers are short, there are no other parties, and nothing urgent is flagged. The record comes back with two items in missing_or_unclear: the customer's previous name appears with two spellings across the two papers they photographed, and the date of the original certificate is missing. You ask the customer to look at the originals and tell you which spelling is correct. The assembly then runs without any further AI step beyond the consistency check, because the form is a one-pager.

The lesson is proportion. Simple files need little tooling, and the value is in the sixty seconds of "are these two spellings the same person?" that a tired human might skip. For complicated files you stop at the red-flag list, not at the tool's limit.

What to do when the AI output looks wrong

Three failure patterns turn up in practice. First, the model invents a plausible value to fill a blank, usually a date or a middle initial; the fix is the "leave null and list it" instruction plus the paper check. Second, it silently tidies a name, turning "Mohammed" into "Muhammad" or dropping a hyphen; the fix is "do not correct names" plus a letter-for-letter name check against ID. Third, it adds a friendly comment such as "this looks like a strong case"; delete the sentence, note the date, and tighten the prompt so output is JSON only. If any of these appear more than once a week, stop and review the stored prompt rather than patching each file.

Step five: the red-flag list and the hand-off script

Some intakes should stop the flow, no matter how smooth the tooling is. Write your own list with your regulator's guidance beside you. A reasonable starting point:

• A hearing or deadline in the next fourteen days, or a deadline you cannot read from the paper.
• Any mention of safety, threats, abuse or a protective order.
• Disputes about children, immigration status, criminal charges or a business's ownership.
• Papers that name several different courts or agencies, or that the customer cannot produce.
• A customer who asks "what should I say?" or "will I win?".

When a flag fires, do not carry on assembling. Use a short script: "I can prepare paperwork you have chosen, but this situation needs advice from a lawyer or a legal-aid service, and I would not want you to lose time. Here is how to find one." Log the flag and the referral in the record. AI can help here only by reading the intake and listing which flags apply; the decision to stop is yours.

Step six: the human review before anything leaves

Review is the control that matters most. A second person, or you after a break, checks each pack against a printed list:

1. Every name matches the customer's ID and papers, letter for letter.
2. The form version is the current one in your library.
3. Every date and amount traces to a document the customer supplied.
4. No box contains words the customer did not give you, other than approved help text.
5. The customer's own choice of document is recorded in the file.
6. Signatures and declarations are left for the customer to complete themselves.
7. The customer has a written note that you are not a lawyer and where to get advice.

Tick the list, initial it and keep it in the file. If a regulator or a customer later asks what happened, a dated checklist is far stronger than your memory.

A worked example with numbers

Say a one-person preparer handles around 30 customers a month, mostly routine court and agency paperwork. Their manual process takes about 65 minutes of clerical time per customer: 20 minutes transcribing intake, 25 minutes filling forms, 15 minutes checking, 5 minutes writing a cover note. After building the flow, transcription falls to about 6 minutes (read the JSON against the papers), form filling to 8 minutes by merge, checking stays at 15 because it is the control, and the cover note to 3 minutes from a template. Total: about 32 minutes. Across 30 customers that is roughly 16 hours saved a month. These are illustrative figures from an imagined practice, not a promise: your forms and customers will differ.

The cost side is modest. ChatGPT Business and Claude Team both list standard seats at $20 a month billed annually or $25 billed monthly, with a two-seat minimum, so a one-person preparer either pays for two seats or uses an individual plan with the model-training switch turned off, plus a hard rule about codes instead of names in prompts. Check both pricing pages before you buy, because seat prices and plan names change. Setup is the expensive part: expect a weekend to build the intake form, mapping sheets and the stored prompts for your first three or four forms.

Data protection and storage

You will hold addresses, finances and often children's details. A few habits cover most of the risk:

• Use customer codes in prompts, with the code-to-name key stored separately.
• Choose a plan whose terms say your content is not used for model training, and turn off chat history where the tool allows.
• Keep the AI step to text you have already minimised: no ID scans, no full bank statements.
• Set a retention period, tell the customer, and delete on schedule.
• If you serve people in the EU, the GDPR applies, so write the lawful basis and the tool's role into your privacy notice.

Telling customers what you do and do not do

Because searchers often arrive wanting to understand the role, publish a short page or handout. A good one has four parts: what a preparer is, what you will do (prepare forms from the customer's answers, check for blanks, explain the form's own wording), what you will not do (advise, choose forms, appear for them, interpret the law), and how to find a lawyer or free legal help. Add one line on AI: "We use an AI tool to tidy typed answers and check forms for mistakes. A person checks everything." Plain wording builds more trust than a list of disclaimers.

For how law firms qualify enquiries before a lawyer is involved, see AI client intake for law firms. For the mechanics of automating consent forms and welcome emails, automate client onboarding with AI covers the plumbing, and can AI fill in forms and supplier portals for you? explains where form-filling tools stumble. If you would rather map your own intake form with someone, that is what my AI implementation consultation is for.

How to check the flow is working

After a month, sample five files and look for three things: the percentage of forms returned or corrected by a court or agency, the number of intake records that needed a second call to the customer, and any instance of AI text appearing in a document without being approved. Zero for the third is the target. If a form is bounced, trace it back to the step that failed (intake, mapping, template currency or review) and fix that step, not the person.

Questions preparers ask about AI and the advice line

Is a legal document preparer allowed to use AI at all?

Usually yes for administrative work such as tidying intake answers, spotting blanks and formatting forms. Rules on what a preparer may do differ by jurisdiction, so check your own licence conditions or regulator guidance, and treat any AI output that selects or explains the law as off limits.

Can AI tell a customer which form they need?

Do not let it. Choosing a form for a person's situation is often exactly what the advice boundary protects. The customer, or a lawyer they consult, decides what to file. Your intake can record the customer's own stated choice, and your tool can check that the answers match that form.

Should I tell customers I use AI?

Yes, in your intake terms, in plain words: what is processed, by which tool, whether it is stored, and that a person checks the output. Customers handing over family, money and housing details should not have to discover the tool later.

Which AI tool is safest for customer data?

Prefer a business plan or an API setup whose terms say your content is not used for training, switch off history where you can, and keep names out of prompts by using codes. Check the vendor's current privacy page, because settings and plans change.

What should a customer do if their situation is complicated?

Your intake should have a red-flag list, such as urgent court dates, disputed facts, children or safety concerns, and a short script that points the customer to a lawyer or legal-aid service. Stopping at that point protects both of you.

Further reads

• How to Turn Handwritten Forms Into Spreadsheet Data With AI — For customers who bring paper answers.
• How to Train ChatGPT on Your Business Information Safely — Keeping your own templates and rules in a project.
• How Small Law Firms Use AI to Draft Letters and Routine Documents — How a law firm handles routine drafting.
• Which Legal Tasks Should a Small Firm Never Hand to AI? — A firm's list of tasks AI should never take.
• AI Literacy Requirements: What Your Staff Need to Know — What anyone touching your AI tools should know.
• Personal Trainer Intake Forms: Build One With AI, Protect Health Data — Another intake form that holds sensitive data.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: a court-run legal document preparer certification programme's public page, OpenAI's ChatGPT Business pricing and business-data pages, Claude's pricing page, and OpenAI's and Anthropic's consumer training-setting help pages (checked 4 November 2026).